This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TMC.2014.2381232, IEEE Transactions on Mobile Computing

IEEE TRANSACTION ON MOBILE COMPUTING, FEB 2014 1

Accurate and Efficient Object Tracking based

on Passive RFID
Lei Yang, Jiannong Cao, Senior Member, IEEE,
Weiping Zhu, Member, IEEE, and Shaojie Tang, Member, IEEE

Abstract—RFID technology has been widely used for object tracking in indoor environment due to their low cost and convenience
for deployment. In this paper we consider RFID reader tracking which refers to continuously locating a mobile object by attaching
it with a RFID reader that communicates with passive RFID tags deployed in the environment. One difficulty is that the RFID
readings gathered from the environment are often noisy. Existing approaches for tracking with noisy RFID readings are mostly
based on using Particle Filter (PF). However, continuous execution of PF has extremely high computational cost, and may be
difficult to be done on mostly resource constrained mobile RFID devices. In this paper, we propose a hybrid method which
combines PF with Weighted Centroid Localization (WCL) to achieve high accuracy and low computational cost. Our observation
is that WCL has the same accuracy with PF with much lower cost if the object’s velocity is low. Our method has two critical
features. The first feature is adaptive switching between using WCL and PF based on the estimated velocity of the mobile object.
The second feature is the further reduction of computational cost by offloading costly PF algorithm onto nearby servers. We
evaluate the performance of our method through extensive simulations and experiments in two real world applications, namely,
indoor wheelchair navigation and in-station Light Rail Vehicle (LRV) tracking at one of Hong Kong MTR depots. The result shows
that our proposed approach has significantly less computational cost than existing PF based methods, while being as accurate
as them.

Index Terms—RFID tracking; particle filter; computation offloading.

1 I NTRODUCTION reader

As one of the most challenging problems in the

area of mobile services, localization has been ex-
tensively studied over the past decade. Compared
with existing solutions using laser scanners, cameras,
or ultrasound, Radio Frequency Identification (RFID)
technique has attracted a lot of interest in recent years
due to its wide adoption by the industry. According
to the types of the used tags, the techniques are
classified into two categories, active tags based tech-
nique [7][15][16][19] and passive tags based technique
[3][4][6][11][12][13][14][17]. Passive tags are more at-
tractive than active tags because of its low cost and
convenience for large-scale deployments.
In this paper we study the problem of location
tracking using passive RFID tags. For convenience,
the tags refer to the passive RFID tags throughout
this paper. RFID based location tracking can be clas-
sified into tag tracking and reader tracking [7]. In tag
tracking as illustrated in Fig.1(a), the object to be
localized is attached with a tag. The RFID readers
• Lei Yang and Jiannong Cao are with the Department of Computing,
Hong Kong Polytechnic University, Hong Kong.
E-mail: csleiyang@comp.polyu.edu.hk, csjcao@comp.polyu.edu.hk
• Weiping Zhu is with the International School of Software, Wuhan
University, China.
E-mail: cswpzhu@gmail.com
• Shaojie Tang is with the Naveen Jindal School of Management, Uni-
versity of Texas at Dallas, U.S.
E-mail: shaojie.tang@utdallas.edu
tag
(a) tag tracking (b) reader tracking
Fig. 1. RFID tracking
are deployed in the environment. As the object moves
in the environment, the readers collect the data. The
readers can either deliver the data to a centralized
server that calculates the position, or cooperate with
each other to estimate the position by themselves. The
positioning result is then returned to the object. In
reader tracking as illustrated in Fig.1(b), each object
to be located carries a RFID reader as well as an
antenna integrated with the reader. The RFID tags are
deployed in the environment. The reader obtains the
data and calculates its own position. Compared with
tag tracking, reader tracking reduces infrastructure
cost by using cheap tags instead of expensive readers.
We focus on reader tracking in this paper.
In the RFID reader tracking,we need to solve a few
challenges. First, we focus on the tracking of mobile
objects rather than localization of stationary objects. It
is more challenging because the algorithm to estimate
the object’s current location must be executed before
a deadline in order to meet the accuracy requirement.

(c) 2015 Crown Copyright. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TMC.2014.2381232, IEEE Transactions on Mobile Computing
IEEE TRANSACTION ON MOBILE COMPUTING, FEB 2014
Second, the RFID readings gathered from real world
are noisy. It means that each tag in the reading range
of the RFID reader is not certainly but possibly to be
detected. Detection failure of tags is normal. The de-
tection failures happen frequently and normally, while
in other localization systems, e.g., infrared based sys-
tem, ultrasonic based systems and WiFi based sys-
tems, the loss of signals from sensors/beacons did
not happen normally. The challenge is to develop
new approaches that can tolerate high uncertainty
of the RFID readings. Third, to tolerate the detec-
tion failures in passive RFID, an effective approach
is to repeat the scanning periodically with a high
frequency. Redundant high-speed scanning of tags
can improve the accuracy under uncertain sensory
readings, but it causes high computational cost, which
may exceed the processing capability of the resource
constrained mobile devices. Thus, the challenge is to
design low computational cost methods that could be
accommodated by the mobile RFID devices.
Existing methods for tracking mobile objects using
noisy RFID readings are often based on Kalman filter
[19], and Particle Filter (PF) [3][6][13][14], also known
as Sequential Monte Carlo Method [20]. Kalman filter
is suitable to be used in linear system with Gaus-
sian noise. However, in many practical RFID tracking
systems, the location evolves none-linearly over time,
and the noise may not be Gaussian. In these none-
linear none Gaussian systems, particle filter can get
more accurate solution than Kalman filter. In particle
filter, the object’s position is represented by a set of
weighted samples (called particles). Two procedures,
sampling and filtering, are executed repeatedly. Dur-
ing the sampling, the particles are sampled randomly
from an area which is usually predicted according
to the motion model of the mobile object. During
the filtering, each particle is first assigned a weight
according to the sensing model. The particles with
low weight are then filtered out. However, continuous
execution of particle filter suffers from high compu-
tational cost on a resource constrained RFID enabled
device. Other existing approaches [1] [2] [10] such as
Weighted Centroid Localization (WCL) are cheap in
computational cost, but yield bad accuracy specially
when the object’s speed is high.
In this paper, we propose a hybrid method for
tracking mobile objects with high accuracy and low
computational cost. This method particularly uses two
critical techniques. First, it adaptively chooses costly
PF and cheap WCL according to the estimated speed
of the object. If the mobile object’s velocity is less
than a threshold, WCL is used; otherwise, the particle
filter is used. This is because of our observation that,
when the object’s speed is low, WCL is able to achieve
at least the same accuracy with particle filter while
costing much less computation. Oppositely, when the
object’s speed is high, the accuracy of WCL is much
worse than particle filter. Furthermore, the motion
(c) 2015 Crown Copyright. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
2
pattern of the moving object (e.g. orientation) becomes
stable under high velocity. This stable pattern could
be taken advantage of to improve the sampling ef-
ficiency, if particle filter is used, and thus to reduce
the computational cost. The complementary property
of PF and WCL in terms of accuracy and efficiency
motivates us to integrate them together.
Second, our hybrid method can offload
intensive computations required by PF onto
nearby infrastructures or servers based on the
quality of network connections. We evaluate our
solution through real world experiments, and show
that our proposed approach with computation
offloading outperforms other approaches in terms
of both accuracy and efficiency. We summarize the
contribution of this paper as follows:
• We propose a hybrid method for object tracking
using noisy passive RFID readings. Compared
with existing particle filter method, it saves a lot
of computational cost while achieving the same
accuracy.
• We design a fast and cheap method to estimate
mobile object’s velocity from the RFID readings.
It is not necessarily applied in our proposed hy-
brid method for object tracking, but can be used
separately in other application scenarios which
require the measurement of object’s speed.
• Besides extensive simulations, we have validat-
ed the method in two real systems, indoor
wheelchair tracking and LRVs (Light Rail Vehi-
cles) tracking in one of Hong Kong MTR depots.
2 M ODELS
2.1 System Model
The RFID reader based tracking system contains three
component: RFID reader, a number of passive RFID
tags, and the environment infrastructures, which are
shown in Fig.1(b). The reader is attached on the
object to be tracked. The infrastructures include the
servers deployed in the area where the object is
moving. The RFID reader can communicate with the
server through wireless networks such as WLAN.
We assume that the RFID reader is integrated with
a circularly polarized antenna. Fig.2 illustrates the
geometrical relationship of the reader and tags. The
reader has a height of H above the ground, and
a circular reading range tr. The number of passive
tags deployed in the tracking area are K. Each tag’s
location, denoted as Ti , is stored in advance at the
reader in association with its tag ID. Whenever the
reader detects a tag, it can access the tag’s location
through the tag ID. The target moves with a velocity
of v. The RFID readings are obtained in a frequency
of f .
It is assumed that detection failure could happen
even when the tag appears in the reader’s reading
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TMC.2014.2381232, IEEE Transactions on Mobile Computing

IEEE TRANSACTION ON MOBILE COMPUTING, FEB 2014 3

Fig. 2. System Model

TABLE 1
Summary of terms and their definitions
Fig. 3. Detection count and sensor model measured
Term Definition
by experiments.
p(r) RFID sensor model
tr Reading range of RFID reader
f Observation frequency of the RFID reader
All tags that lie beyond the distance tr have a
M The number of tags
Ti The location of i-th tag
probability of 0 to be detected and all tags lying
vth Threshold of the velocity
within the range tr has a detection probability that
Nth Threshold of the maximum duration exponentially decreases as the distance r increases.
rki Detection status of tag i at time k.If detect, λ, α and β are parameters that measure detection
rki = 1; otherwise,rki = 0 probability when a target is within a distance of r.
zk The RFID reading at time k, zk = {rki }M i=1
We conduct an experiment to obtain the sensor
lk Location of target at time k model in a classroom with enough open space. The
βi The spatial weight of tag i RFID device is RRU1861, an UHF RFID integrated
λt (k) Temporal weight in Extended WCL. k is with a circularly polarized antenna. The reader is
historical time, t is current time placed at a height of 35cm from the floor while the
tags are uniformly distributed onto the floor, which is
partitioned into 11 × 11 grids with each tag occupying
range. (Actually this failure always happens in real a grid. The size of grid is 7.5 cm. The detection count
world because of either the hardware failure or wire- of each tag among a total of 2000 scans is recorded.
less environment inference). Let RFID reading at time It is found that distance is much more significant in
k be zk = {rki }M i
i=1 , where rk indicates whether or affecting the successful detection than orientation of
not tag i has been detected at the time k; if it has the tag with respect to the antenna. That is why only
been detected, then rki = 1; otherwise rki = 0. The distance is considered in our model. The measurement
objective is to estimate the continuous location of the is repeated by rotating the antenna a bit each time and
mobile object using the uncertain RFID reading. The the results are averaged over different orientations of
following section describes a RFID sensing model in the antenna with respect to fixed tags. Figure 3 shows
order to capture this uncertainty of RFID reading. a visualized probabilistic sensor model. Through a
model based fitting using the experimental data, we
2.2 RFID Sensing Model also obtain the parameters in Equation (1), α = 2.3374
, β = 0.0011, λ = 0.7921.
Previous works [11][12] in RFID localization assume
that a perfect binary sensor model exists. The reader
is capable of detecting the tags that lie within its
3 T HE H YBRID M ETHOD FOR RFID R EAD -
sensing range and not the tags beyond it. In reality, ER T RACKING
the successful detection of a specific tag depends on 3.1 Overview of Our Method
a large number of parameters: 1) relative position of The objective in this paper is to design an accurate
a tag from the reader; 2) absorbing and reflecting and efficient method for object tracking using the
materials in the environment; 3) noise existing in the passive RFID. Tracking is continuous localization of
environment. These parameters lead to uncertainty of the moving object. The frequency of localization is
successful detection of tags by the reader. To ensure usually determined by the application. Given the
that our algorithm is as true to reality as possible, localization frequency, the accuracy is measured by
we assume a probabilistic sensing model by equation the average localization error over the time slots. The
(1), which is motivated by the existing sensor model efficiency is measured by the average computational
proposed in WSN [5][18]. time of the localization algorithm in one time slot. The
( α computation time should be less than the required
λe−βr , r < tr
p(r) = (1) localization period; otherwise, it can not satisfy the
0, r > tr real time property of the tracking application. The

(c) 2015 Crown Copyright. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TMC.2014.2381232, IEEE Transactions on Mobile Computing
IEEE TRANSACTION ON MOBILE COMPUTING, FEB 2014
Start
Speed Estimation
Network Quality
Algorithm
Esitmation
Selection
Particle Offloading
WCL
Filter Decision
Execution
Local Execution
RFID Device
Fig. 4. Work flow of the hybrid method
less the computation time is, the better efficiency
the algorithm has. Traditional methods in tracking
mainly use two methods, namely Weighted Centroid
Localization (WCL) and Particle Filter (PF). WCL is
cheap but yields bad accuracy in many application
scenario, while PF has good accuracy, but it costs
much computational cost. We aims to design an RFID
tracking method which has better performance than
traditional WCL and PF methods.
Our method includes two simple ideas. In the
following, we briefly explain them and illustrate the
intuition behind them.
•Integration of WCL and PF. We study the accu-
racy of WCL and PF when they are used in RFID
reader tracking. We observe that, on one hand, the
accuracy of WCL depends on the velocity of the
object. If the velocity is relatively slow, WCL has
satisfactory accuracy; otherwise the accuracy becomes
very bad. On the other hand, the accuracy of PF is
much better, and does not change much with the
velocity of moving object. Thus, we ask the question,
if we can estimate the velocity of the moving object,
can we replace the PF by WCL when the velocity is
relatively small in order to save the computation cost,
while not sacrificing the accuracy?
•Computation Offloading. The mobile RFID de-
vice inherently has limited computing capability, and
can cause very high computation cost on the device.
However, as more and more computation power is
integrated into the environmental infrastructures, and
there exists abundant wireless techniques connecting
the RFID device and infrastructures, can we offload
the complex computation involved in PF onto the
infrastructures?
With these two questions, we design a hybrid track-
ing method for better accuracy and efficiency. Fig.4
shows the work flow of our method. At each time
slot, the RFID devices first estimate the speed of the
moving object based on recent RFID readings. After
the speed estimation, the component of algorithm
selection is performed. This component is responsible
for choosing the better algorithm from WCL and PF
according to current speed. We have a threshold of
speed for the algorithm selection. If the speed is
large than the threshold, PF is selected to estimate
(c) 2015 Crown Copyright. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
4
the location in current time slot; otherwise, WCL
is selected. As WCL is very cheap in computation,
if WCL is selected, the algorithm is then executed
locally on the RFID device. However, if the PF is
selected, the algorithm seeks to offload the compu-
Remote
tations onto the nearby servers based on the network
quality. The component of network quality estimation
Server is performed periodically.
3.2 Weighted Centroid Localization (WCL)
The Centroid Localization (CL) is proposed at first in
the outdoor localization for wireless devices [9]. The
device to be localized calculates its own position by
a centroid determination from all the positions of the
beacons in range. However, the CL has two limitations
when it is used in RFID positioning system. First,
as described in the previous section some tags may
be missed even when it is within the reading range.
Second, CL method inherently results in a coarse esti-
mation. For limitation one, the failure of identification
of tags can be avoided by multiple detection attempts.
For limitation two, a weighted CL (WCL) method
[1][2][10] is applied to improve the accuracy.
The key idea of WCL is to assign greater weight
to those tags which are estimated to be closer to
the target and less weight to the farther tags. The
target position is estimated according to the following
equation:
PM
i=1 (βi Ti )
l= P n (2)
i=1 βi
where βi represents the weight assigned to tag i.
The weight should be determined by the distance of
tags from the target according to geometrical rules. In
our method, the weight is simply approximated by
the count of successful detection during a number of
detection attempts. This may be explained by the fact
that the detection probability is inversely proportional
to theP distance. The weight of tag i is given by
βi = k rki .
3.2.1 Extend WCL Method in Time Domain
WCL may result in an increasing error when it is
applied into tracking mobile object. This is because
the target keeps moving when the reader is gathering
data. To overcome the above issue, we extend WCL
method in time domain for tracking a moving ob-
ject. In specific, the current location is estimated by
smoothing RFID readings of recent N observations.
Historical RFID reading are assigned different weights
in estimating current position at time t. Intuitively, the
latest RFID reading should be given largest weight
in this context. For convenience, these weights are
named temporal weights in order to distinguish from
weights introduced previously, which are in space do-
main and called spatial weights. Therefore, the spatial
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TMC.2014.2381232, IEEE Transactions on Mobile Computing
IEEE TRANSACTION ON MOBILE COMPUTING, FEB 2014
weights of tag i at time t are calculated according to
t Trangular Pattern
X 0.3
βi (t) = λt (k)rki ,
k=t−N
where λt (k) is temporal weight of RFID reading at
time k. In our method, temporal weights are deter-
mined by intersection of reading range at time k and
that of time t.
Area(St ∩ Sk )
λt (k) = , k = t − N, t − N + 1, · · · , t − 1, t
Area(St )
(4)
Through some geometric rules, Equation (4) can be
simplified by the following equation:
(t − k)v
λt (k) = f [ ],
2f · tr
where
( √
2[arccos α − (1 − α) 2α − α2 ]/π, 0 < α < 1;
f (α) =
0, α >= 1
(6)
v is the current average moving speed. f is the
observation frequency of RFID reader. tr presents the
reading range of the reader. Equation (5)(6) shows
that, given f and tr, the temporal weight is a decreas-
ing function in terms of the time difference (t − k)
and the speed v. The reason by including v into
this function is that the faster a object is moving the
sooner previous readings will expire. Then the next
challenging problem is how to estimate the average
moving speed of the object?
3.2.2 How to estimate the object’s speed
In this work, we first calculate the time duration di (N )
that each tag i stays in the reader’s reading range in
the last N time slots (assume N > di ∀i). Then we
select the maximum one dmax (N ) = max di to roughly
estimate average moving speed as follows:
2tr · c
v=
dmax (N )
In Equation (7), 2tr is the diameter of the reading
range and 0 < c ≤ 1 is scaling constant depending on
the tag density. Above speed calculation is motivated
by the following observation: the tag with largest di
tends to be the one which is closest to the object’s ac-
tual moving path. Theoretically, if a tag is right on the
moving path, it should stay in reader’s reading range
until the object traveling distance 2tr. The probability
that the closest tag is right on the object’s moving path
is clearly increased as the density of the tags increases.
Therefore c is an increasing function of tag density.
(c) 2015 Crown Copyright. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
5
0.4
Location Error (divided by tr)
0.35
Square Pattern
(3) 0.25
0.2
0.15
0.1
0.05
0
0.2 0.4 0.6 0.8 1 1.2
Tag density (divided by tr)
1.4
Fig. 5. The location error of WCL depending on the
tags density
3.2.3 Considerations for Transmission Range tr and
Tag Density
Most RFID readers in current market are able to adjust
their power levels. Assume the reader antenna was
(5)
circularly polarized, the transmission range tr usually
increases as the power level increases. The larger tr is,
the more tags will be included into the centroid esti-
mation of the object each time WCL is performed. The
question is what the optimal tr is when using WCL
method in object tracking. Intuitively, the optimal tr
depends on the tag density. Another question is how
the tag density affects the accuracy of WCL. Is it true
that the more densely the tags are deployed, the more
accurate results will be obtained through WCL?
We analyze WCL by numerical simulations featur-
ing a variety of tag densities. The tags are uniformly
distributed onto a two-dimension plane according to
a regular pattern. Only triangular pattern and square
pattern are considered. The tag density is defined as
the size of spacing between tags. Spacing in triangular
pattern is normalized into that of square pattern,
following the rule that the node number per square
meter remains the same. The transmission range of
the reader tr is preset to a constant value. The speed
of the moving object is also preset to a constant,
satisfying v = 2tr·f
60 . The tag density increases from
0.2tr to a upper bound of 1.4tr, beyond which there
exists blind zone where the object detects no tag. The
sensing model of Equation (1) is used to generate the
simulated RFID readings. The location error is defined
as the distance between the exact position and the
(7) estimated position of the reader.
Fig.5 shows that the location error depends on
the tags density. We can have the following conclu-
sions from Fig.5. First, the optimal density for WCL
algorithm is located in the interval of (0.8tr, 0.9tr).
Increasing tag density under the spacing size of 0.8tr
nearly does not improve the accuracy. However, a
sparser density upper bound the spacing size of 0.9tr
leads to serious degeneracy of the accuracy. Second,
the necessary tr of the RFID reader should be as
large as 1.1a, where a is the average spacing size
between two tags, indicating the deployment density
of tags. Third, the deployment pattern of tags, either
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TMC.2014.2381232, IEEE Transactions on Mobile Computing
IEEE TRANSACTION ON MOBILE COMPUTING, FEB 2014
square pattern or triangular patter, does not affect the
accuracy.
3.3 Particle Filtering
In this subsection, we solve this object tracking prob-
lem through a Bayesian approach [8]. Using this
approach, two models are generally required: first, a
model describing the evolution of the object’s location
with time (the system model); second, a model relat-
ing the noisy RFID readings (the measurement mod-
el). These models are presented in a probabilistic form.
Considering the evolution of the position sequence of
a moving object, lk = (xk , yk , θk ), k = 1, 2, 3, · · ·, we
have the following system model
v
xk = xk−1 + cos θk−1 (8)
f
v
yk = yk−1 + sin θk−1 (9)
f
θk = θk−1 + U (−∆θ, ∆θ)
where xk , yk are two coordinates of the object’s
location, θk is the heading orientation of the objec-
t, U (−∆θ, ∆θ) is an increase angle of the heading
direction which yields a uniform distribution, v is
the velocity of the object, and f is the observation
frequency of the reader. The measurement model is
given by
M
Y upper bound of f , if the object moves in a high
p(zk |lk ) = p(rki |lk ) (11)
i=1
p(rki |lk ) = rki p(di ) + (1 − rki )(1 − p(di )),
where di is the distance of tag i from target and p(di )
is the RFID sensor model described in Section 2. The
objective is to recursively estimate lk from the RFID
readings, denoted as zk = {rki }Mi=1 .
From a Bayesian perspective, the tracking problem
is to recursively calculate some degree of belief in the
state lk at time k, given the data z1:k up to time k. It
is required to construct the posterior pdf , usually in
two stages: prediction and update.
A particle filter approximates the posterior pdf at
time k by a set of M particles. Each particle contains
a position hypothesis lki = (xik , yki , θki ) and a weight
wki . wki indicates the importance of the i-th particle. A
position estimate lˆk can be obtained by lˆk = i=1 wki lki
PM
. The particle filtering algorithm takes the following
steps in a recursive fashion:
• Prediction The target position at time k is predi-
cated by propagating all particle positions at time
k − 1 according to system model, Equation(8-10).
lki ∼ p(lk |lk−1
i
, v, ∆θ)
We assume that v has a Gaussian distribution,
N (vk , σk2 ). The model probabilistically describes
the transition of system from old state for a given
measurement vk .
(c) 2015 Crown Copyright. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
6
• Updating Particles are reweighted according to
the latest RFID reading zk and the measurement
model, Equation(11-12).
wki = wk−1
i
p(zk |lki )
The weight is then normalized by wki =
PM
wki / i=1 wki .
• Resampling A new set of n particles with e-
qual weights 1/n is obtained from the old set of
particles. The possibility of choosing particle i is
proportional to its weight wki . The objective of
resampling is to reduce the effects of degeneracy.
The basic idea is to eliminate the particles that
have small weights and to concentrate on parti-
cles with large weights.
3.4 Adaptive Algorithm Selection of WCL and PF
3.4.1 Limitation of WCL
WCL faces two weaknesses in object tracking. First, if
(10)
the tags are deployed sparsely onto the ground, it is
possible that tr can not achieve the necessary value
1.1a (see the discussion above about Fig.5), even when
the Reader’s power reaches the maximum level. In
this case, the accuracy of WCL will be degenerated
seriously, as shown by Fig.5. Second, the observation
frequency of RFID f always has a up-bound in order
not to affect the detection rate of tags. Given this
speed, most of the measurements in the past become
invalidate to calculate the current location. Therefore,
(12) the accuracy will decrease as the speed of the object
increases, which is also demonstrated by Fig.7 in
Section 4.
3.4.2 Limitation of Particle Filter
Although Particle Filter (PF) has been widely used
in object tracking for its satisfactory accuracy, the
biggest limitation of PF comes from its high compu-
tational cost. Table 3 in Section 4 demonstrates its
costliness in computation. Since mobile objects are
usually equipped with resource constrained devices,
continuously running PF on these devices may not be
possible.
3.4.3 The hybrid method to integrate WCL with PF
Aware of limitations of both WCL and PF, the ques-
tion here is, can we combine these two methods
together in order to achieve both high accuracy and
efficiency in object tracking? To realize this objective,
our strategy is that in any case WCL does not have
worse accuracy than PF, WCL will be adopted to
replace PF in object tracking. These cases are formu-
lated by the following conditions: trmax >= 1.1a and
v <= vth = 2tr·f
Nth , where vth is the velocity threshold,
Nth is the required duration (denoted as the number
of time slots) by WCL that a tag can stay in the reading
range of a moving object.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TMC.2014.2381232, IEEE Transactions on Mobile Computing
IEEE TRANSACTION ON MOBILE COMPUTING, FEB 2014
Usually with the known density of deployed tags,
the first condition, trmax >= 1.1a, can be easily
judged before the runtime. If this condition is satis-
fied, the power of the reader will be tuned to the
level, in which case tr = 1.1a. During the runtime,
the moving speed of the object is estimated according
to Equation (7). If the estimated speed is less than
vth , WCL will be executed; otherwise, PF will be per-
formed. Instead of deciding which method should be
performed before the runtime, we integrate WCL with
PF by an adaptive execution of these two methods
in the runtime according to the estimated speed of
the moving object. This complies with most practical
situations that the object moves with a varying speed
rather than a constant speed.
Note that Nth can affect a lot the performance of
the hybrid method. If Nth is set too high, that means
the threshold of the speed vth is too low, the hybrid
method has little improvement on the computational
cost compared with particle filter. Oppositely, if Nth is
set too low, the hybrid method sacrifices the accuracy
too much. It is hard to prove theoretically the optimal
value of Nth . However, we can acquire an optimal
value of Nth from the numerical simulations. Section
4 demonstrates an optimal value of Nth in a simulated
scenario.
3.5 Computation Offloading
To further reduce the computation complexity on
the RFID device, our method can offload the
computation-intensive part of the tracking algorithm
to a nearby server. Although offloading computation
onto the server can lower the computational cost on
the RFID device, it brings additional communication
cost due to the transmission of required data between
the RFID device and the server. The principle of the
offloading decision is that, given the algorithm func-
tion, if reduced computation cost by remote execution
is large than the incurred communication cost, then
it is better to offload this function onto the server;
otherwise the function is better to be executed locally.
•Offloading Principle. Assume the estimation of
location at time t with the RFID readings zt needs
time τl if it is performed locally on the RFID device,
and needs time τr if it is executed remotely on the
server. Let ds denote the size of transmission data
in remote execution. Let B denote the bandwidth of
the wireless connection between the RFID device and
the server. We have the following offloading decision.
If τm − τr > dBs , the algorithm is then offloaded to
the server at this time k; otherwise, the algorithm is
executed on the RFID device. In the following, we
describe how to collect τr , τl , ds and B in real systems.
In the hybrid method, the WCL and PF algorithm-
s are adaptively executed according to the object’s
velocity. We collect both the local and remote com-
putation cost of WCL and PF from our experimen-
tal implementation, which is described in details in
(c) 2015 Crown Copyright. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
7
TABLE 2
Local and remote computation cost
Execution time τl τr
WCL 2.7 ms 0.12 ms
Particle filter 215 ms 6.8 ms
RFID Readings at time t
Particles state and
result at time t
RFID Readings at time t+1
Particles state and
result at time t+1
...
RFID Device Server
Fig. 6. The data transmission between RFID device
and server if particle filter is executed remotely
Section 5. The main processing platform of the RFID
device in our implementation is the Mica2 sensor
mote. The remote server is the Fujitsu LifeBook S
Series. Table 2 shows the one time computational cost
of WCL and PF respectively in location and remote
execution. The execution of WCL at both the mobile
device and the server causes very little computation
cost. For simplicity in our system, we only consider
the offloading of PF. Note that the execution time of
PF varies depending on the number of particles. In
Table 2, the execution time of PF is measured when
the number of particles is 2000.
To make the optimal offloading decision, we need
to estimate the communication cost of PF when it is
executed remotely. At each time of location estima-
tion, the PF needs to predict the state of the particles
according to the mobility model of the object, and
then update the weights of the particles based on the
RFID readings in this time slot. Fig.6 shows the data
transmission between the RFID device and the server
when PF is performed remotely. At the beginning of
the time slot t, the RFID device uploads the RFID
readings zt to the server. The server does the predic-
tion and resampling, and returns the state of particles
as well as the estimated location at time t. The data
required to transfer on the network mainly includes
the observed RFID readings zt , the state of particles,
i.e., the parameters wki and lki . The reason that the
state of particles need to be returned to the RFID
device is that the wireless network connection may be
intermittent. We need to backup the state on the RFID
device. In case the network connection is down, the
PF can be continued locally with the backup particles
state. Note that the size of transmission data increases
as the number of particles in PF increases.
With the size of transmission data, we need to mea-
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TMC.2014.2381232, IEEE Transactions on Mobile Computing

IEEE TRANSACTION ON MOBILE COMPUTING, FEB 2014 8

0.2
TABLE 3
PF (Particle number = 2000) Efficiency Comparison between WCL and PF
0.18 PF (Particle number = 1000)
PF (Particle number = 500)
0.16
WCL v(m/s) 0.1 0.2 0.3 0.4 0.5 0.6
Location Error (m)

0.14
WCL(ms) 119 86 75 64 65 66
0.12
PF-2000(ms) 7593 8484 10568 9792 9260 9276
0.1
PF-1000(ms) 4654 4927 6012 4873 4825 5103
0.08

0.06
PF-500(ms) 2769 2876 3421 2785 2631 2724

0.04
0 0.1 0.2
Velocity (m/s)
Fig. 7. Accuracy comparison between WCL and PF
sure the bandwidth of the wireless network, in order
to estimate the communication cost. In our method,
we periodically estimate the network bandwidth by
transferring a 10K file between the device and server.
The bandwidth is the testing file size divided by the
transmission time. In our experiment, the RFID con-
nects the server through ZigBee motes. We measure
the bandwidth fluctuation through carrying with the
RFID device and walking around in one classroom.
We found that the bandwidth indeed fluctuates. The
fluctuation period is on the order of several seconds.
4 P ERFORMANCE E VALUATION
S IMULATIONS
In this section we evaluate the performance of the
proposed hybrid method. We consider the following
metrics: location error (accuracy) and computational
cost (efficiency). First, we evaluate respectively and
compare the performance of WCL and Particle Filter
(PF). The parameter we consider in the comparison is
velocity because it is the main factor that distinguishes
the performance of these two methods. Then, we eval-
uate how much the hybrid method can improve the
performance by considering a trajectory along which
the object moves with a varying speed. We further
evaluate the impact of the computation offloading
strategy onto the performance.
4.1 Performance Comparison between WCL and
Particle Filter
The simulated environment is a 4m × 4m plane. The
RFID tags are densely deployed onto the plane. The
spacing size of tags a is 0.27m. The reading range
of the reader tr is 0.3m. The observation frequency
f is 10Hz. Assume that the object moves with a
constant speed around a rectangle trajectory in the
2-Dimension plane. The RFID model described in
Section 2.2 is used to generate the simulated RFID
readings. The location is estimated continuously dur-
ing one trajectory through different tracking methods.
The accuracy is represented by an average of the
location error at each time slot.
0.3 0.4 0.5 0.6
Fig.7 compares the location error of WCL and PF
under different velocities. We plot the location error
of PF when the number of particles are respectively
2000, 1000, and 500. We can see that the location error
of WCL increases as the velocity increases while the
accuracy of particle filter has no association with the
velocity. PF has better accuracy than WCL when the
velocity exceeds 0.2m/s. When the velocity is less
than the value of 0.2m/s, WCL has almost the same
accuracy with PF. Fig.7 also shows that the accuracy of
PF becomes better as the number of particles increas-
es. Table 3 compares the execution time of the two
methods. Simulations for both algorithms are run on
the same platform (Matlab Environment on a Fujitsu
LifeBook S Series). To compare the computational
THROUGH efficiency among the cases of different velocity, the
number of trajectories are tuned for each velocity so
that the tracking time in all the cases keeps the same.
From Table 3, we can see that WCL is obviously more
efficient than PF whatever the target’s velocity is. The
PF has more computational cost as the number of
particles increases.
Fig.8a and Fig.8b show the location error at each
time slot as well as the estimated trajectory when the
object’s velocity is 0.6m/s. We can see that the location
error of WCL is larger than that of PF. It is interesting
to note that the location deviation of WCL is along
the trajectory while the location deviation of PF is
perpendicular to the trajectory. This is because that
executing the WCL periodically for object tracking has
a critical limitation. If no tag is detected in current
time slot, the tags detected in the past time slot
will dominate the centroid estimation of the current
position. Therefore, the result at current time is very
likely to be the real location in the past time slot. In
specific, this location error becomes very serious when
the object’s speed is high.
4.2 The Hybrid Method of WCL and PF
We compare the performance of the hybrid method
with that of other two methods, WCL and Particle
Filter, by considering a more practical scenario where
the object’s velocity varies during the trajectory. Fig.8c
shows a simple case that the object circulates around
the rectangle with two different velocities alternative-
ly. Around four corners the object moves with a slow

(c) 2015 Crown Copyright. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TMC.2014.2381232, IEEE Transactions on Mobile Computing
IEEE TRANSACTION ON MOBILE COMPUTING, FEB 2014
(a) WCL
Fig. 8. (a)(b) visualize the location error at each time slot as well as the entire estimated trajectory when the
object’s velocity keeps constant at v = 0.6m/s; (c) is a simulated trajectory with a varying velocity
TABLE 4
Performance comparison between the three methods
Schemes WCL PF HybridMethod
LocationError(m) 0.187 0.068 0.062
ExecutionTime(s) 0.08 4.53 1.26
velocity, v = 0.15m/s. In other intervals the object
runs with a fast velocity, v = 0.6m/s.
How to select the threshold vth and Nth when the
hybrid method is performed in the scenario? It has
been shown in Fig.7 that the optimal vth is 0.2m/s.
Below this threshold, WCL has the same accuracy
with PF, but above it WCL has much worse accuracy
than PF. The Nth is computed by Nth = 2tr·f
vth = 30.
Table 4 presents the location error and execution
time of the three methods. Note that the number of
particles is set as 2000 in all PF executions. It shows
that hybrid method leads to much lower location error
than WCL. Moreover, compared with particle filter,
the hybrid method saves about 75% computational
cost while achieving the same accuracy.
4.3 The Offloading Strategy
In this section, we evaluate the influence of offloading
strategy onto performance, particularly in term of
the efficiency. We compare the performance of the
following five tracking methods.
•WCL-Local. In this method, the tracking algorith-
m is WCL. The algorithm is executed on the RFID
device.
•PF-Local. In this method, the tracking algorithm is
Particle Filter (PF). The algorithm is executed on the
RFID device.
•PF-Offloading. In this method, the tracking algo-
rithm is Particle Filter (PF). The execution of PF can
be offloaded onto the server at each time slot.
•Hybrid-Local. In this method, the tracking al-
gorithm is adaptively selected from WCL and PF
according to the estimated velocity. The algorithm is
executed on the RFID device.
(c) 2015 Crown Copyright. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
9
(b) Particle Filter (c) A Trajectory
200
Bandwidth (KB/s)
High power
150 Low power
100
50
0
0 20 40 60 80 100
Time (s)
Fig. 9. Network bandwidth trace
•Hybrid-Local. In this method, the tracking algo-
rithm is adaptively selected from WCL and PF. The
algorithm can be offloaded onto the server.
In this evaluation, the object’s moving trajectory
and velocity have the same setting with Fig.8c. The
8 8
time used in this trajectory is 0.15 + 0.6 = 66.7s. As the
sampling frequency is 10 Hz, we have about 667 time
slots in the whole trajectory. The number of particles
in all PF executions is set as 2000. Table 2 shows the
profiling computational cost of WCL and PF in each
step (or time slot) respectively on the RFID device and
the server. Since WCL runs very fast on local device,
we do not consider the offloading of WCL. When PF
is offloaded onto the server, we measure that the size
of transmitted data between the RFID device and the
service is 5.2 kB.
We collect real network bandwidth traces from one
classroom. In the classroom, one notebook is deployed
as the server. The notebook is connected with a TelosB
mote. The RFID device attached with Mica2 mote
communicates with the server through ZigBee. We
carry the device and move randomly around the
classroom. The network bandwidth is measured by
transferring one 10kB testing file. The measurement
is done every 5 seconds. Fig.9 shows two network
traces when the transmission power of the ZigBee
motes is set with various levels. In high power setting,
both motes have level 256 in transmission power,
while in low power setting, both motes have level 128
in transmission power. We use the network trace in
high power setting to evaluate the performance with
offloading strategy. Note that only the first 66.7s of
the whole network trace is used in the evaluation.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TMC.2014.2381232, IEEE Transactions on Mobile Computing

IEEE TRANSACTION ON MOBILE COMPUTING, FEB 2014 10

250
250
200
Efficiency (ms)

Algorithm Running Time (ms)

150 200
Hybrid−Local
100 Hybrid−Offloading
50 150 PF−Local
PF−Offloading
0
WCL PF−Local PF−Offloading Hybrid−Local Hybrid−Offloading
100
(a) Efficiency/Algorithm running time
0.5
Accuracy/Location Error (m)

50
0.4

0.3 0
0 50 100 150 200 250 300
0.2 Wireless bandwidth (KB/s)
0.1

0 Fig. 11. Algorithm running time varies depending on

WCL PF−Local PF−Offloading Hybrid−Local Hybrid−Offloading
the network bandwidth
(b) Accuracy/Location error
Fig. 10. Comparison of various methods in terms of
accuracy and efficiency

Fig.10a compares the efficiency of the five tracking

methods under this evaluation setting. We can see
that both the PF and the hybrid method with the
offloading strategy have better efficiency than the cor-
responding methods without offloading. In particular,
the PF-Offloading needs 42.3 ms running time on av-
(a) (b)
erage at each time slot, while the PF-Local needs 215
ms running time. The offloading strategy brings about Fig. 12. The deployment of RFID system for indoor
5X efficiency improvement to the PF. The Hybrid- wheelchair navigation
Offloading needs 7.9 ms running time, and is more
6m
than 5X better than Hybrid-Local that has 45.2 ms 0.5m
running time per time slot. Among the five tracking 8
methods, WCL and Hybrid-Offloading have much 1
less execution cost than other three methods. Taking 9
the accuracy into account, as shown in Fig.10b, the
4m

3 4
Hybrid-Offloading has the best performance among
5 7
the five methods. 6
10
We then evaluate how the computational efficiency
of the methods changes depending on the band- 2
width of wireless connection between the RFID de-
vice and the server. Fig.11 plots the efficiency as the
bandwidth increases from zero to 300kB. The two
methods with offloading strategy, Hybrid-Offloading Fig. 13. Wheelchair trajectories in a 4m × 6m class-
and PF-Offloading, have decreasing execution cost as room
the bandwidth increases. This is because under the
offloading strategy, large bandwidth leads to short the tags deployment. The reader contains 18 discrete
time in the data transmission between the device power levels. We tune the power on level 12 so that
and server. When the bandwidth is as little as zero, the reading range tr is as far as 60cm, satisfying the
the methods with offloading strategy have the same equation tr >= 1.1a, where a = 50cm.
execution cost with the methods running locally on The wheelchair can be controlled with a joystick
the device. controller. We attach an extension unit to the joystick
for controlling the output signal. This unit is connect-
5 E XPERIMENTS ed to a Mica2 Mote through serial port. The speed of
the wheelchair is controlled by a pair of output volt-
5.1 Indoor wheelchair tracking age values. These voltage values can be set through
In the scenario of wheelchair tracking, the RFID tags the Mica2 Mote. In our experiment, the wheelchair
are densely deployed on the ground in a 4m × 6m moves along a line with a varying speed. During
classroom. The spacing size of tags (density) is 50cm. the first half of the entire trajectory, the wheelchair’s
The RFID device we attached on the chassis of the speed is set to 0.1m/s. In the second half, the speed
wheelchair is an integrated UHF RFID reader with a is increased to 0.6m/s. The observation frequency f
circularly polarized antenna. Fig.12 and Fig.13 shows is 10Hz. By selecting Nth = 40, the threshold of

(c) 2015 Crown Copyright. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TMC.2014.2381232, IEEE Transactions on Mobile Computing
IEEE TRANSACTION ON MOBILE COMPUTING, FEB 2014
TABLE 5
Experiment results for wheelchair tracking
Methods WCL PF-O
Location error (m) 0.482 0.186
Time per step (ms) 3.6 89.2
0.9
0.8
0.7
0.6
CDF (%)
0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4
Location Error (m)
Fig. 14. CDF of tracking errors
the speed for the hybrid method is set to vth =
2tr·f
Nth = 0.3m/s. In the classroom, we deploy a laptop
as the server. The laptop is connected with a TelosB
mote through the USB, such that the Mica2 mote on
the RFID reader can offload computations onto the
server through ZigBee protocol. The bandwidth of
the ZigBee connection between the RFID reader and
server is measured every 5 seconds.
To evaluate the location error, the real location
is calculated through the integral of the preset
speed. The location is computed periodically us-
ing the four tracking methods, WCL-Local (WCL),
PL-Offloading (PL-O), Hybrid-Local (Hybrid-L) and
Hybrid-Offloading (Hybrid-O). For each method, we
repeat the experiment 10 times by choosing different
trajectories in the classroom. Fig.13 shows the 10
trajectories generated in the 4m×6m classroom. The
result is averaged over the 10 experiments. Table 5
presents the average location error and execution time
per step of the three methods. Fig.14 shows the CDF of
tracking errors under the four algorithms. The hybrid
method (Hybrid-O) almost achieve the same accuracy
as PF (PF-O), but outperforms PF a lot in term of
computational cost.
5.2 LRV tracking in Hong Kong MTR depot
In the scenario of LRV tracking, we installed an RFID
reader below the LRV, shown in Fig.15. Note that
our RRU1861 URF RFID reader/attenna does not
have computation and networking capability. To use
it in location estimation, we usually need to attach
it with external devices. In the wheelchair tracking
experiment, we attach it with the Mica2 mote, while in
this experiment, we attach it with one Sharp SH631W
smart phone. We deploy our self-made mesh routers
(c) 2015 Crown Copyright. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
11
Hybrid-O Hybrid-L
0.195 0.193
19.4 95.7
(a) (b)
Fig. 15. The deployment of RFID system at one MTR
depot
[20] along the rail such that all the testing area are
covered by the mesh network. The location can be
WCL either calculated locally by the smart phone which is
PF−O
Hybrid−O
physically connect with the RFID reader, or calculated
Hybrid−L by the server on the mesh networks.
0.5
0.6 0.7 0.8
We deploy a total of 42 RFID tags along the track.
When the LRV is moving along the tagged rail, the
RFID reader keeps querying for the tags. The obser-
vation frequency of the reader f is 20Hz. We have
tested two tracks. On one track, the train is moving
at the speed of 15km/h. On the other track, the train
is moving at 35km/h. We adjust the power level of
the reader to the maximum level, level 18, so that the
reading range, tr = 0.8m, can satisfy the condition
required by WCL, which is tr > 1.1a, where a is about
0.7m.
The duration Nd (measured by the number of
time slots) that a tag is able to stay in the reading
range of the reader is respectively 3 and 7 in case
of v = 35km/h and v = 15km/h. For one tag, if
all the N detections fail, this tag will definitely fail
to be detected; if any one of the Nd detections is
successful, this tag will be detected successfully. We
found that the detection rate of tags in the two tests
are quite different. The detection rate is about 70%
when v = 15km/h. When v = 35km/h, the detection
rate decreases to 45%.
In our experiment, Nth is set to an experience
value, 30. The LVR’s speed in both the two tests are
above the threshold. According to the rules of the
hybrid method, the location should be estimated by
particle filter. In order to evaluate the location error,
the real location is also approximated by the integral
of the LRV’s speed. By performing the PF algorithm
based on the data set, we obtain that the location
error is 0.45m when v = 15km/h. However, when
v increases to 35km/h, the location error increases
to 1.2m. We also evaluate the computational cost of
the PF algorithm in this LRV tracking application. We
compare the computation cost of PF in the case the
algorithm is performed locally on the smart phone
and the case the algorithm is offloaded onto the server
on the mesh network. We find the offloading strategy
reduce the computation cost per time slot from 137
ms to 24.8 ms.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TMC.2014.2381232, IEEE Transactions on Mobile Computing
IEEE TRANSACTION ON MOBILE COMPUTING, FEB 2014
z
z
©
y ¶
y
© of the hexahedron. Fig.16 illustrates the 3-D object
x
¶
x
Fig. 16. RFID based object tracking in 3-D environ-
ment
6 D ISCUSSION
This paper not only provides a hybrid method of
object tracking, but also provides a guide for users
to deploy the RFID based tracking system under
different performance requirements. A tradeoff be-
tween the accuracy and efficiency can be achieved by
dynamically adjusting vth or Nth in the runtime. If
the computational cost is critical in the system, a low
value of Nth is preferred. If high accuracy is desired,
a high value of Nth is chosen.
RFID Sensing Model. In real world, the RFID sens-
ing model will be quite different when the device is
performed in different environments. As the sensing
model in Section 2 is measured in the same configured
classroom, it is able to be used into the experiment of
wheelchair navigation. However, in the LRV tracking
experiment, a more general sensing model, shown
in Equation (13), is applied into the PF algorithm
instead, where p0 is a constant. Only when one tag has
been detected, the step of filtering out the invalidate
particles will be executed according to the sensing
model. In other times, the prediction step based on
the estimated speed will be executed.
( [7][15][19]. In these works, reference tags with known
p0 , r < tr;
p(r) = (13)
0, r > tr
As shown in the LRVs tracking experiment, neither
the PF nor the hybrid method is suitable for tracking
high speed moving objects cases, the tags detection
rate is usually low. ( The detection rate is as low as
45% when the LVR’s speed is v = 35km/h ). The bad
accuracy in this scenario can be explained as follows:
1) when the speed is too high, which also means
the value of Nd is too low, our method to estimate
the speed will not work. Without information of the
speed, the efficiency of particle sampling decreases
seriously. 2) The invalidate particles are not filtered
out until the next tag is successfully detected. Thus,
when the detection rate is low, these unfiltered invalid
particles will leads to large location error.
3-D Object Tracking Model. The proposed method
can be easily extended into 3-D scenes. We consider
(c) 2015 Crown Copyright. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
12
the targets move in a hexahedron, e.g., a warehouse,
a reading room, etc. The passive tags with known
positions are deployed on the floor and ceil. The tags’
locations are represented by (x, y, z) three coordinates.
The z coordinate of tags on the floor is zero, and
the z coordinate of tags on the ceil is the height
tracking environment which has been discussed in
our previous work [28]. We assume that the reading
range of the RFID reader is sphere. The RFID sensing
model in 3-D scenes has the same formulation with
Equation (1), where the r represents the 3-D distance
between the tag and the reader. The system model
which represents the evolution of the target’s location
with time should change from Equation (8-10) to the
Equation (14-18). With the new 3-D system model,
the algorithms discussed in our paper for 2-D object
tracking can be also applied in the 3-D environment.
v
xk = xk−1 + cos θk−1 cos φk−1 (14)
f
v
yk = yk−1 + cos θk−1 sin φk−1 (15)
f
v
zk = zk−1 + sin θk−1 (16)
f
φk = φk−1 + U (−∆φ, ∆φ) (17)
θk = θk−1 + U (−∆θ, ∆θ) (18)
7 R ELATED W ORKS
RFID Localization/Tracking. There exists a variety of
approaches to RFID-based localization. Many early
systems use active tags, because the active tag is
like the traditional sensors, being able to transmit
signals, and thus the previous Radio Frequency (RF)
based localization approaches, such as WLAN and
WSN, can be easily applied into these RFID systems
positions are deployed regularly on the covered area.
Lionel et al. [7] present LANDMARC that uses kNN
technique for localizing unknown active tags. Huang
et al. [15] utilizes Bayesian inference to calculate the
target position. Yamano et al. [19] use support vector
machine to estimate the robot’s location. Our work in
this paper focuses on the object tracking using passive
RFID tags, due to its cheapness and widely usage in
industry.
Most related works on localization and/or objective
tracking based passive RFID tags use the geometrical
and range free methods to estimate the location of
moving objects. Han et al. [11] present localization
scheme for an indoor mobile robot using passive
RFID tags. A triangular pattern of arranging the RFID
tags on the floor is adopted to reduce the estimation
error of the conventional square pattern. Park et al.
[12] propose a method using read-time of IC tags to
reduce the localization error of an RFID navigation
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TMC.2014.2381232, IEEE Transactions on Mobile Computing
IEEE TRANSACTION ON MOBILE COMPUTING, FEB 2014
system. However, both of these two literatures assume
a binary sensor model and do not consider detection
failure of passive tags.
The literatures [3][6][13][14] consider a probabilistic
sensor model for the RFID reader, and use particle
filter to estimate the location. They focus on the
tracking of robot, and assume that the robot has
abundant resources to run the costly particle filter
method. However, in this paper, we study more
general scenario where the objects to be tracked are
usually attached with cheap embedded RFID devices.
Running PF on these devices may cause large latency,
and thus can not satisfy the real time property of the
tracking application.
Jeffrey et.al [27] evaluates the performance of parti-
cle filter algorithm running on various mobile devices,
and compares particle filter with other deterministic
localization methods. They focus on applying particle
filter in two localization systems, namely infrared
badge system from VersusTech and MIT Cricket ul-
trasound system. Both the two localization systems
are using active sensors/beacons. In our work, we
focus on the localization with passive sensors/ RFID
tag. The environment is deployed with passive RFID
tags, which can only receive signals passively from
the objects. The communications between passive tags
and readers is more prone to failures. Furthermore,
the collisions among multiple tags could increase the
probability of detection failures. The existing methods
like centroid based methods and particle filter can
not directly applied to solve the localization problem
based passive RFID.
Phase value has been used to localize or track
RFID tags to achieve high accuracy in recent works
[29][30][31]. Pinit [29] compares the multipath profile
(phase related information) between the target tag and
the reference tags, and estimate the targets location
from its nearest references tags. Unlike existing RSS
based methods, Pinit can achieve high accuracy in
multipath or non-line-of-sight environments. Howev-
er, it needs relatively dense deployment of reference
tags to guarantee high accuracy. Real time tracking of
objects may suffering from latency problem due to the
time in collection of tags multipath profiles. BagPos
[30] is anchor free, and does not need extra informa-
tion from the reader or tags. However, it sacrifices the
area of feasible region. Tagoram [31] utilizes the phase
value of the backscattered signal to track the mobile
RFID tags in real time, and can achieve centimeter-
level accuracy in real world deployments.
Related Works on Computation Offloading. Com-
putation offloading has been widely used to solve
the resource poverty problem of mobile devices. Nor-
mally to achieve a fine granularity for computation
offloading, a partitioning scheme has been proposed
by researchers [24] [25] [26]. Given that the application
is composed of a set of dependent tasks, computation
partitioning is to decide whether each task is executed
(c) 2015 Crown Copyright. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
13
on the mobile device or on the cloud, such that the
cost is minimized. Related works on computation
partitioning considers the application modeling and
optimization of various objectives. They differ in the
cost models, which consider either one of these factors
such as energy consumption of the mobile device,
application delay, and data transmission amount on
the network, or the combination of these factors.
Most related works on computation offloading fo-
cus on the modeling, optimization and the design of
runtime system to support the offloading. They do
not evaluate to what extend computation offloading
method can improve the efficiency of RFID tracking
applications. Meanwhile, to maximize the profit of
computation offloading, we believe that the design
of offloading strategy should be considered together
with the algorithm design of RFID tracking applica-
tions.
8 C ONCLUSION
In this paper we have studied the problem of tracking
mobile objects that are able to detect the presence of
passive RFID tags when traveling in their proximity.
We proposed a hybrid method for achieving high
accuracy and efficiency in object tracking. Through
numerical evaluations, we showed that our method
is more computational efficient than PF while guar-
anteing the same accuracy with PF. We also imple-
mented two real RFID based tracking systems using
our proposed method in two applications, namely, the
wheelchair navigation and LRVs tracking at one of
Hong Kong MTR depots. For the former application,
our method can achieve better performance than both
PF and WCL in either efficiency or accuracy. For
the latter application, in most cases our method can
achieve good accuracy, but in some case the LRV
moves in a very high speed our method do not
perform well due to the low detection rate of the RFID
tags.
ACKNOWLEDGMENTS
The research is partially supported by Hong Kong
RGC under GRF Grant 510412, and the National High-
Technology Research and Development Program (863
Program) of China under Grant 2013AA01A212.
R EFERENCES
[1] R. Behnke, and D. Timmermann. AWCL: adaptive weighted
centroid localization as an efficient improvement of coarse
grained localization. In Proc. Int. Workshop on Positioning, Navi-
gation and Communication 2008, pages 243C250.
[2] J. Blumenthal, R. Grossmann, F. Golatowski, and D. Timmer-
mann. Weighted Centroid Localization in Zigbee-based Sensor
Networks. In Proc. IEEE Int. Symposium on Intelligent Signal
Processing 2007, pages 1C6.
[3] D. Hahnel, W. Burgard, D. Fox, K. Fishkin, and M. Philipose.
Mapping and localization with RFID technology. In Proc. IEEE
Int. Conf. on Robotics and Automation 2004, pages 1015C1021.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TMC.2014.2381232, IEEE Transactions on Mobile Computing
IEEE TRANSACTION ON MOBILE COMPUTING, FEB 2014
[4] D. Joho, C. Plagemann, and W. Burgard. Modeling RFID Signal
Strength and Tag Detection for Localization and Mapping. In
Proc. IEEE Int. Conf. on Robotics and Automation 2009, pages 3160-
3165.
[5] A. Ghosh, and S. Das. Coverage and connectivity issues in wire-
less sensor networks: A survey. Pervasive and Mobile Computing,
Vol.4, Issue 3, pages 303C334, June 2008.
[6] E. Giampaolo, and F. Martinelli. Robot localization by sparse
and passive RFID tags. In Proc. IEEE Int. Symposium on Industrial
Electronics 2010, pages 1937C1942.
[7] L. Ni, Y. Liu, Y. Lau, and A. Patil. LANDMARC: indoor location
sensing using active RFID. In Proc. IEEE Int. Conf. on Pervasive
Computing and Communications 2003, pages 407C415.
[8] M. Arulampalam, S. Maskell, N. Gordon, and T. Clapp. A
tutorial on particle filters for online nonlinear/non-gaussian
bayesian tracking. IEEE Tran. on Signal Processing, Vol.50, Issue
2, pages 174C188, 2002.
[9] N. Bulusu, J. Heidemann, and D. Estrin. Gps-less low cost
outdoor localization for very small devices. IEEE Personal Com-
munications Magazine, Vol.7, Issue 5, pages 28C34, 2000.
[10] N. Bulusu, J. Heidemann, and D. Estrin. Centroid localization
of uncooperative nodes in wireless networks using a relative
span weighting method. EURASIP Journal on Wireless Commu-
nications and Networking, Vol.10, Issue 11, pages 1C11, 2010.
[11] S. Han, H. Lim and J. Lee. An efficient localization scheme
for a differential-drive mobile robot based on rfid system. IEEE
Trans. On Industrial Electronics, Vol.54, Issue 6, pages 362C369,
2007.
[12] S. Park and S. Hashimoto. Autonomous mobile robot naviga-
tion using passive rfid in indoor environment. IEEE Trans. on
Industrial Electronics, Vol.56, Issue 7, pages 2366C2373, 2009.
[13] S. Schneegans, and P. Vorst. Using RFID snapshots for mobile
robot self-localization. In Proc. European Conference on Mobile
Robots 2007, pages 1C6.
[14] P. Vorst, and A. Zell. Semi-autonomous learning of an RFID
sensor model for mobile robot self-localization. In Proc. Euro-
pean Robotics Symposium 2008, pages 273C282.
[15] X. Huang, R. Janaswamy, and A. Ganz. Scout: Outdoor lo-
calization using active RFID technology. In Proc. Int. Conf. on
Broadband Communications, Networks and Systems 2006, pages
1C10.
[16] X. Liu, M. Corner, and P. Shenoy. Ferret: RFID localization for
pervasive multimedia. In Proc. ACM Int. Conf. on Ubiquitious
Computing 2006, pages 422C440.
[17] Y. Zou and K. Chakrabarty. Sensor deployment and target
localization in distributed sensor networks. IEEE Trans. on
Embedded Computing and System, Vol.3, Issue 1, pages 61C91,
2004.
[18] K. Yamano, K. Tanaka, M. Hirayama, and et al. Self-
localization of mobile robots with RFID system by using sup-
port vector machine. In Proc. IEEE/RSJ Int. Conf. on Intelligent
Robots and Systems 2004, pages 3756C3761.
[19] M. Richard, and S. Nozer. Understanding the Kalman Filtr. The
American Statistician, Vol.37, Issue 2, 1983
[20] J. Cao, K. Xie, W. Wu, C. Liu, and et al. HAWK: Real-world
Implementation of High-performance Heterogeneous Wireless
Network for Internet Access. In Proc. of Int. Workshop on Next
Generation Network Architectures 2009, pages 214-220.
[21] Z. Li, C. Wang, and R. Xu. Computation offloading to save
energy on handheld devices: a partition scheme. In Proc. of
Int. Conf. on Compilers, Architecture, and Synthesis for Embedded
Systems 2011, pages 238-246.
[22] K. Kumar, and Y. Lu. Cloud computing for mobile users:
Can offloading computation save energy. IEEE Computer, Vol.43,
Issue 4, pages 51-56, 2009.
[23] M. Satyanarayanan, and P. Bahl, R. Caceres, and N. Davies.
Cloud computing for mobile users: Can offloading computation
save energy. IEEE Pervasive Computing, Vol.6, Issue 4, pages 12-
23, 2009.
[24] L. Yang, J. Cao, S. Tang, T. Li, and A. Chan. A framework
for partitioning and execution of data stream applications in
mobile cloud computing. In Proc. of IEEE Int. Conf. on Cloud
Computing 2012, pages 794-782.
[25] E. Cuervoy, A. Balasubramanianz, and D. Cho. MAUI: making
smartphones last longer with code offload. In Proc. of ACM Int.
Conf. on Mobile Systems, Applicaitons and Services 2010, pages
49-62.
(c) 2015 Crown Copyright. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
14
[26] B. Chun, S. Ihm, P. Maniatis, M. Naik, and A. Patti.
CloneCloud: elastic execution between mobile device and
cloud. In Proc. of Sixth Conf. on Computer Systems 2011, pages
301-314.
[27] J. Hightower, and G. Borriello. Particle Filters for Location
Estimation in Ubiquitous Computing: A Case Study. In Proc.
of Int. Conf. on Ubicomp 2004.
[28] W. Zhu, J. Cao, Y. Xu, L. Yang, and J. Kong. Fault-Tolerant
RFID Reader Localization Based on Passive RFID Tags. IEEE
Trans. on Parallel and Distributed Systems, Vol.25, Issue 8, pages
2065 - 2076, 2014.
[29] J. Wang and D. Katabi. Dude, Wheres My Card? RFID Po-
sitioning that Works with Multipath and Non-line-of-sight. in
Proc. of ACM SIGCOMM, 2013.
[30] T. Liu, L. Yang, Q. Lin, Y. Guo, Y. Liu. Anchor-free Backscatter
Positioning for RFID Tags with High Accuracy. in Proc. of IEEE
INFOCOM, 2014.
[31] L. Yang, Y. Chen, X. Li, C. Xiao, M. Li, Y. Liu. Tagoram: Real-
Time Tracking of Mobile RFID Tags to High Precision Using
COTS Devices. in Proc. of ACM MobiCom, 2014.
Lei Yang received the BSc degree in elec-
tronic engineering from Wuhan University,
P.R.China, in 2007, and the MSc degree in
computer science from Institute of Comput-
ing Technology, Chinese Academy of Sci-
ence, P.R.China, in 2010. He is a post-
doctoral fellow of the Department of Comput-
ing at Hong Kong Polytechnic University. His
research interest includes mobile cloud com-
puting, RFID systems, and social computing.
Jiannong Cao received the BSc degree from
Nanjing University, China, in 1982, and the
MSc and PhD degrees from Washington S-
tate University, USA, in 1986 and 1990, all
in computer science. He is currently a chair
professor and the head of the Department
of Computing at Hong Kong Polytechnic Uni-
versity. His research interests include paral-
lel and distributed computing, computer net-
works, mobile and pervasive computing, fault
tolerance, and middle-ware.
Weiping Zhu received the BSc degree from
Chongqing University, China, in 2003, the
MSc degree from Shanghai JiaoTong Uni-
versity, China, in 2006, and the PhD de-
gree from Hong Kong Polytechnic University,
Hong Kong, in 2013, all in computer science.
He is an associate professor in the Interna-
tional School of Software at Wuhan Universi-
ty. His research interest includes RFID, wire-
less sensor networks, distributed computing,
and pervasive computing.
Shao-Jie Tang received his Ph.D degree
from Department of Computer Science at
Illinois Institute of Technology in 2012. He re-
ceived B.S. in Radio Engineering from South-
east University, P.R. China in 2006. He is
an assistant professor with Naveen Jindal
School of Management, University of Texas
at Dallas. His main research interests focus
on wireless networks (including sensor net-
works and cognitive radio networks), social
networks, pervasive computing, mobile cloud
computing and algorithm analysis and design.